Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fairness without Demographics through Learning Graph of Gradients

Yingtao Luo, Zhixun Li, Qiang Liu, Jun Zhu

TL;DR

This work tackles fairness without access to sensitive demographic data by exploiting the information contained in last-layer gradients. It introduces the Graph of Gradients (GoG), a soft grouping mechanism that connects samples via $K$-nearest gradient neighbors and uses an adversarial network to weight samples during training, optimizing a Rawlsian, worst-off objective with minimal accuracy sacrifice. The authors provide theoretical justifications that gradient signals better capture demographics than input features and demonstrate, across COMPAS, BNP, and MIMIC-III, that GoG achieves superior worst-group fairness and robustness to label noise with competitive overall accuracy. The approach is scalable (linear complexity) and interpretable, offering a privacy-preserving, effective solution for real-world fairness challenges in diverse domains.

Abstract

Machine learning systems are notoriously prone to biased predictions about certain demographic groups, leading to algorithmic fairness issues. Due to privacy concerns and data quality problems, some demographic information may not be available in the training data and the complex interaction of different demographics can lead to a lot of unknown minority subpopulations, which all limit the applicability of group fairness. Many existing works on fairness without demographics assume the correlation between groups and features. However, we argue that the model gradients are also valuable for fairness without demographics. In this paper, we show that the correlation between gradients and groups can help identify and improve group fairness. With an adversarial weighting architecture, we construct a graph where samples with similar gradients are connected and learn the weights of different samples from it. Unlike the surrogate grouping methods that cluster groups from features and labels as proxy sensitive attribute, our method leverages the graph structure as a soft grouping mechanism, which is much more robust to noises. The results show that our method is robust to noise and can improve fairness significantly without decreasing the overall accuracy too much.

Fairness without Demographics through Learning Graph of Gradients

TL;DR

This work tackles fairness without access to sensitive demographic data by exploiting the information contained in last-layer gradients. It introduces the Graph of Gradients (GoG), a soft grouping mechanism that connects samples via -nearest gradient neighbors and uses an adversarial network to weight samples during training, optimizing a Rawlsian, worst-off objective with minimal accuracy sacrifice. The authors provide theoretical justifications that gradient signals better capture demographics than input features and demonstrate, across COMPAS, BNP, and MIMIC-III, that GoG achieves superior worst-group fairness and robustness to label noise with competitive overall accuracy. The approach is scalable (linear complexity) and interpretable, offering a privacy-preserving, effective solution for real-world fairness challenges in diverse domains.

Abstract

Machine learning systems are notoriously prone to biased predictions about certain demographic groups, leading to algorithmic fairness issues. Due to privacy concerns and data quality problems, some demographic information may not be available in the training data and the complex interaction of different demographics can lead to a lot of unknown minority subpopulations, which all limit the applicability of group fairness. Many existing works on fairness without demographics assume the correlation between groups and features. However, we argue that the model gradients are also valuable for fairness without demographics. In this paper, we show that the correlation between gradients and groups can help identify and improve group fairness. With an adversarial weighting architecture, we construct a graph where samples with similar gradients are connected and learn the weights of different samples from it. Unlike the surrogate grouping methods that cluster groups from features and labels as proxy sensitive attribute, our method leverages the graph structure as a soft grouping mechanism, which is much more robust to noises. The results show that our method is robust to noise and can improve fairness significantly without decreasing the overall accuracy too much.

Paper Structure

This paper contains 16 sections, 3 theorems, 11 equations, 3 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Group fairness for classification
  4. Fairness without demographics
  5. Theoretical Analysis
  6. Problem Formulation
  7. Definition of gradient
  8. Relationship between gradients and sensitive demographics
  9. Methodology
  10. Experiments
  11. Experimental Setup
  12. Performance Comparison
  13. Robustness Study
  14. Ablation Study
  15. Interpretability Study
  16. ...and 1 more sections

Key Result

theorem 1

The mutual dependence between gradients (the combination of input features and model error) and sensitive attributes is larger than the mutual dependence between input features and sensitive attributes. If we denote sensitive demographics as s, model prediction error as U, and input features as x, I

Figures (3)

  • Figure 1: The proposed Graph of Gradients algorithm framework.
  • Figure 2: Robustness study of different fairness algorithms across three datasets, evaluated by the Worst-Group AUC/Accuracy.
  • Figure 3: The distribution of input features and last-layer gradients upon different race groups for MIMIC-III dataset (Purple: 'White', Yellow: 'BLACK', Green: 'HISPANIC', Other races are not considered for convenience).

Theorems & Definitions (3)

  • theorem 1
  • Lemma 1
  • proposition 1